Daily Arxiv

전 세계에서 발간되는 인공지능 관련 논문을 정리하는 페이지 입니다.
본 페이지는 Google Gemini를 활용해 요약 정리하며, 비영리로 운영 됩니다.
논문에 대한 저작권은 저자 및 해당 기관에 있으며, 공유 시 출처만 명기하면 됩니다.

Sekai: A Video Dataset towards World Exploration

Federated Learning for MRI-based BrainAGE: a multicenter study on post-stroke functional outcome prediction

One-Step Diffusion for Detail-Rich and Temporally Consistent Video Super-Resolution

Capturing Polysemanticity with PRISM: A Multi-Concept Feature Description Framework

Efficient Retail Video Annotation: A Robust Key Frame Generation Approach for Product and Customer Interaction Analysis

WebXAII: an open-source web framework to study human-XAI interaction

Refining music sample identification with a self-supervised graph neural network

Decoupled Classifier-Free Guidance for Counterfactual Diffusion Models

Essential-Web v1.0: 24T tokens of organized web data

Adaptive Guidance Accelerates Reinforcement Learning of Reasoning Models

LeVERB: Humanoid Whole-Body Control with Latent Vision-Language Instruction

Screen Hijack: Visual Poisoning of VLM Agents in Mobile Environments

SP-VLA: A Joint Model Scheduling and Token Pruning Approach for VLA Model Acceleration

Serving Large Language Models on Huawei CloudMatrix384

Machine Learning Methods for Small Data and Upstream Bioprocessing Applications: A Comprehensive Review

Med-U1: Incentivizing Unified Medical Reasoning in LLMs via Large-scale Reinforcement Learning

Two Heads Are Better than One: Simulating Large Transformers with Small Ones

BreastDCEDL: Curating a Comprehensive DCE-MRI Dataset and developing a Transformer Implementation for Breast Cancer Treatment Response Prediction

Semantic Preprocessing for LLM-based Malware Analysis

A Minimalist Method for Fine-tuning Text-to-Image Diffusion Models

Human-like Forgetting Curves in Deep Neural Networks

Convergent Linear Representations of Emergent Misalignment

LearnAlign: Reasoning Data Selection for Reinforcement Learning in Large Language Models Based on Improved Gradient Alignment

TARDIS STRIDE: A Spatio-Temporal Road Image Dataset and World Model for Autonomy

Autonomous Computer Vision Development with Agentic AI

The Memory Paradox: Why Our Brains Need Knowledge in an Age of AI

SWE-Factory: Your Automated Factory for Issue Resolution Training Data and Evaluation Benchmarks

Using Language and Road Manuals to Inform Map Reconstruction for Autonomous Driving

Reinforcing Spatial Reasoning in Vision-Language Models with Interwoven Thinking and Visual Drawing

PlantBert: An Open Source Language Model for Plant Science

Info-Coevolution: An Efficient Framework for Data Model Coevolution

SDE-SQL: Enhancing Text-to-SQL Generation in Large Language Models via Self-Driven Exploration with SQL Probes

SafeGenBench: A Benchmark Framework for Security Vulnerability Detection in LLM-Generated Code

UniWorld-V1: High-Resolution Semantic Encoders for Unified Visual Understanding and Generation

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

GraphRAG-Bench: Challenging Domain-Specific Reasoning for Evaluating Graph Retrieval-Augmented Generation

Towards Efficient Few-shot Graph Neural Architecture Search via Partitioning Gradient Contribution

Optimizing Sensory Neurons: Nonlinear Attention Mechanisms for Accelerated Convergence in Permutation-Invariant Neural Networks for Reinforcement Learning

CryoCCD: Conditional Cycle-consistent Diffusion with Biophysical Modeling for Cryo-EM Synthesis

More Thinking, Less Seeing? Assessing Amplified Hallucination in Multimodal Reasoning Models

Calibrating Pre-trained Language Classifiers on LLM-generated Noisy Labels via Iterative Refinement

Dynamic Risk Assessments for Offensive Cybersecurity Agents

SSR-Zero: Simple Self-Rewarding Reinforcement Learning for Machine Translation

Robo2VLM: Visual Question Answering from Large-Scale In-the-Wild Robot Manipulation Datasets

Breaking the Compression Ceiling: Data-Free Pipeline for Ultra-Efficient Delta Compression

Representation Learning with Mutual Influence of Modalities for Node Classification in Multi-Modal Heterogeneous Networks

Learning Dynamics in Continual Pre-Training for Large Language Models

Mask-PINNs: Regulating Feature Distributions in Physics-Informed Neural Networks

Assessing Tenstorrent's RISC-V MatMul Acceleration Capabilities

SPIN-ODE: Stiff Physics-Informed Neural ODE for Chemical Reaction Rate Estimation

PerceptionLM: Open-Access Data and Models for Detailed Visual Understanding

DeepSelective: Interpretable Prognosis Prediction via Feature Selection and Compression in EHR Data

Boosting multi-demographic federated learning for chest radiograph analysis using general-purpose self-supervised representations

AerialVG: A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations

PR-Attack: Coordinated Prompt-RAG Attacks on Retrieval-Augmented Generation in Large Language Models via Bilevel Optimization

TALE: A Tool-Augmented Framework for Reference-Free Evaluation of Large Language Models

Decentralized Collective World Model for Emergent Communication and Coordination

RL2Grid: Benchmarking Reinforcement Learning in Power Grid Operations

Efficient but Vulnerable: Benchmarking and Defending LLM Batch Prompting Attack

LLMs in Disease Diagnosis: A Comparative Study of DeepSeek-R1 and O3 Mini Across Chronic Health Conditions

Can We Detect Failures Without Failure Data? Uncertainty-Aware Runtime Failure Detection for Imitation Learning Policies

QG-SMS: Enhancing Test Item Analysis via Student Modeling and Simulation

Dynamic Knowledge Integration for Evidence-Driven Counter-Argument Generation with Large Language Models

AlignDistil: Token-Level Language Model Alignment as Adaptive Policy Distillation

Eau De $Q$-Network: Adaptive Distillation of Neural Networks in Deep Reinforcement Learning

Hierarchical and Modular Network on Non-prehensile Manipulation in General Environments

Selective Use of Yannakakis' Algorithm to Improve Query Performance: Machine Learning to the Rescue

FRIDA to the Rescue! Analyzing Synthetic Data Effectiveness in Object-Based Common Sense Reasoning for Disaster Response

Batayan: A Filipino NLP benchmark for evaluating Large Language Models

From RAG to Memory: Non-Parametric Continual Learning for Large Language Models

Conformal Inference under High-Dimensional Covariate Shifts via Likelihood-Ratio Regularization

ShapeLib: Designing a library of programmatic 3D shape abstractions with Large Language Models

Rethinking Fine-Tuning when Scaling Test-Time Compute: Limiting Confidence Improves Mathematical Reasoning

Guaranteed prediction sets for functional surrogate models

FDLLM: A Dedicated Detector for Black-Box LLMs Fingerprinting

Each Rank Could be an Expert: Single-Ranked Mixture of Experts LoRA for Multi-Task Learning

MonoSOWA: Scalable monocular 3D Object detector Without human Annotations

Representation Learning of Point Cloud Upsampling in Global and Local Inputs

Multi-Preference Optimization: Generalizing DPO via Set-Level Contrasts

Incivility and Rigidity: The Risks of Fine-Tuning LLMs for Political Argumentation

Song Form-aware Full-Song Text-to-Lyrics Generation with Multi-Level Granularity Syllable Count Control

Learning Multi-Branch Cooperation for Enhanced Click-Through Rate Prediction at Taobao

On the Limits of Language Generation: Trade-Offs Between Hallucination and Mode Collapse

Web Archives Metadata Generation with GPT-4o: Challenges and Insights

Cyclic Vision-Language Manipulator: Towards Reliable and Fine-Grained Image Interpretation for Automated Report Generation

A Implies B: Circuit Analysis in LLMs for Propositional Logical Reasoning

Adapting While Learning: Grounding LLMs for Scientific Problems with Intelligent Tool Usage Adaptation

FALCON: Feedback-driven Adaptive Long/short-term memory reinforced Coding Optimization system

Efficient Mixture-of-Expert for Video-based Driver State and Physiological Multi-task Estimation in Conditional Autonomous Driving

ALTA: Compiler-Based Analysis of Transformers

Learning to Route LLMs with Confidence Tokens

Core Knowledge Deficits in Multi-Modal Language Models

AssistantX: An LLM-Powered Proactive Assistant in Collaborative Human-Populated Environment

MaPPER: Multimodal Prior-guided Parameter Efficient Tuning for Referring Expression Comprehension

Can Large Language Models Replace Human Subjects? A Large-Scale Replication of Scenario-Based Experiments in Psychology and Management

LogProber: Disentangling confidence from contamination in LLM responses

V2X-VLM: End-to-End V2X Cooperative Autonomous Driving Through Large Vision-Language Models

xGen-MM (BLIP-3): A Family of Open Large Multimodal Models

Understanding and Reducing the Class-Dependent Effects of Data Augmentation with A Two-Player Game Approach

PromptDSI: Prompt-based Rehearsal-free Instance-wise Incremental Learning for Document Retrieval

LlamaDuo: LLMOps Pipeline for Seamless Migration from Service LLMs to Small-Scale Local LLMs

Created by

Haebom

저자

Chansung Park, Juyong Jiang, Fan Wang, Sayak Paul, Jing Tang

개요

본 논문은 클라우드 기반의 독점적 대규모 언어 모델(LLM) 사용의 운영 의존성, 개인 정보 보호 문제, 지속적인 인터넷 연결 필요성과 같은 과제를 해결하기 위해, 서비스 지향적 LLM의 지식과 능력을 더 작고 로컬에서 관리 가능한 모델로 원활하게 마이그레이션하는 LLMOps 파이프라인인 "LlamaDuo"를 제시합니다. LlamaDuo는 서비스 LLM을 사용하여 생성된 합성 데이터셋을 사용하여 소규모 언어 모델을 미세 조정합니다. 미세 조정된 모델의 성능이 기대에 미치지 못하면, 서비스 LLM이 생성한 추가적인 유사 데이터를 사용하여 추가 미세 조정을 통해 자동으로 개선됩니다. 이러한 다단계 프로세스는 소규모 모델이 특정 다운스트림 작업에서 서비스 LLM의 기능과 일치하거나 능가할 수 있도록 보장하여 제약된 환경에서 AI 배포를 관리하기 위한 실용적이고 확장 가능한 솔루션을 제공합니다. 다양한 최첨단 LLM을 사용한 광범위한 실험을 통해 다양한 다운스트림 작업에서 LlamaDuo의 효과, 적응성 및 경제성을 입증합니다. 파이프라인 구현은 https://github.com/deep-diver/llamaduo 에서 확인할 수 있습니다.

GitHub - deep-diver/llamaduo: This project showcases an LLMOps pipeline that fine-tunes a small-size LLM model to prepare for the outage of the service LLM.

This project showcases an LLMOps pipeline that fine-tunes a small-size LLM model to prepare for the outage of the service LLM. - GitHub - deep-diver/llamaduo: This project showcases an LLMOps pipe...

시사점, 한계점

•

시사점:

◦

클라우드 의존성을 줄이고 로컬 환경에서 LLM 사용을 가능하게 함으로써 운영상의 유연성과 안정성을 향상시킵니다.

◦

개인 정보 보호 문제를 완화하고 데이터 프라이버시를 강화합니다.

◦

인터넷 연결 없이도 LLM 기능을 활용할 수 있도록 오프라인 사용을 지원합니다.

◦

서비스 LLM의 성능을 유지하거나 능가하는 소규모 모델을 효율적으로 생성하는 방법을 제공합니다.

◦

다양한 다운스트림 작업에서 LlamaDuo의 효과성, 적응성 및 경제성을 실험적으로 검증합니다.

•

한계점:

◦

합성 데이터셋의 품질에 따라 성능이 크게 영향을 받을 수 있습니다.

◦

서비스 LLM의 기능을 완벽하게 복제하는 것은 어려울 수 있으며, 특정 작업에 대한 성능 저하가 발생할 수 있습니다.

◦

미세 조정 과정에 상당한 컴퓨팅 자원이 필요할 수 있습니다.

◦

LlamaDuo의 성능은 사용되는 서비스 LLM과 소규모 모델의 특성에 따라 달라질 수 있습니다.

◦

현재 공개된 GitHub 레포지토리의 완성도 및 사용 편의성에 대한 추가적인 검증이 필요합니다.

Made with Slashpage